figure()
ion()
pl = MultilinePlotter(autoscale=1.2, xlim=[0, 50], ylim=[0, 1])
pl.setLineStyle(linewidth=2)
numbExp=25 #number of experiments
for runs in range(numbExp):
# create environment
#Options: Bool(OpenGL), Bool(Realtime simu. while client is connected), ServerIP(default:localhost), Port(default:21560)
env = ShipSteeringEnvironment(False)
# create task
task = GoNorthwardTask(env,maxsteps = 500)
# create controller network
net = buildNetwork(task.outdim, task.indim, outclass=TanhLayer)
# create agent with controller and learner
agent = FiniteDifferenceAgent(net, SPLA())
# learning options
agent.learner.gd.alpha = 0.3 #step size of \mu adaption
agent.learner.gdSig.alpha = 0.15 #step size of \sigma adaption
agent.learner.gd.momentum = 0.0
batch=2 #number of samples per gradient estimate (was: 2; more here due to stochastic setting)
#create experiment
experiment = EpisodicExperiment(task, agent)
prnts=1 #frequency of console output
epis=2000/batch/prnts
#actual roll outs
filename="dataSPLA08NoRew"+repr(int(random.random()*1000000.0))+".dat"
wf = open(filename, 'wb')
for updates in range(epis):
for i in range(prnts):
from pybrain.tools.shortcuts import buildNetwork
from pybrain.rl.environments.cartpole import CartPoleEnvironment, BalanceTask
from pybrain.rl.agents.finitedifference import FiniteDifferenceAgent
from pybrain.rl.learners import SPLA
from pybrain.rl.experiments import EpisodicExperiment
from scipy import random
numbExp = 12
for runs in range(numbExp):
env = CartPoleEnvironment()
# create task
task = BalanceTask(env, 200)
# create controller network
net = buildNetwork(4, 1, bias=False)
# create agent with controller and learner
agent = FiniteDifferenceAgent(net, SPLA())
# learning options
agent.learner.gd.alpha = 0.05
agent.learner.gdSig.alpha = 0.1
agent.learner.gd.momentum = 0.9
agent.learner.epsilon = 6.0
agent.learner.initSigmas()
# agent.learner.rprop = True
experiment = EpisodicExperiment(task, agent)
batch = 16
prnts = 10
epis = 50000 / batch / prnts
save = False
rl = []
for updates in range(epis):
saveNet = False
saveName = "grasp.wgt"
numbExp = 1 #number of experiments
for runs in range(numbExp):
# create environment
#Options: XML-Model, Bool(OpenGL), Bool(Realtime simu. while client is connected), ServerIP(default:localhost), Port(default:21560)
env = CCRLEnvironment()
# create task
task = CCRLGlasVarTask(env)
# create controller network
net = buildNetwork(len(task.getObservation()),
hiddenUnits,
env.actLen,
outclass=TanhLayer)
# create agent with controller and learner
agent = FiniteDifferenceAgent(net, SPLA())
# learning options
agent.learner.gd.alpha = 0.2 #step size of \mu adaption
agent.learner.gdSig.alpha = 0.085 #step size of \sigma adaption
agent.learner.gd.momentum = 0.0
#Loading weights
if loadNet:
agent.learner.original = loadWeights("grasp.wgt")
agent.learner.gd.init(agent.learner.original)
agent.learner.epsilon = 0.2
agent.learner.initSigmas()
batch = 2 #number of samples per gradient estimate
#create experiment
experiment = EpisodicExperiment(task, agent)
from pybrain.tools.shortcuts import buildNetwork
from pybrain.rl.environments.cartpole import CartPoleEnvironment, BalanceTask
from pybrain.rl.agents.finitedifference import FiniteDifferenceAgent
from pybrain.rl.learners import SPLA
from pybrain.rl.experiments import EpisodicExperiment
from scipy import random
numbExp = 12
for runs in range(numbExp):
env = CartPoleEnvironment()
# create task
task = BalanceTask(env, 200)
# create controller network
net = buildNetwork(4, 1, bias=False)
# create agent with controller and learner
agent = FiniteDifferenceAgent(net, SPLA())
# learning options
agent.learner.gd.alpha = 0.05
agent.learner.gdSig.alpha = 0.1
agent.learner.gd.momentum = 0.9
agent.learner.epsilon = 6.0
agent.learner.initSigmas()
# agent.learner.rprop = True
experiment = EpisodicExperiment(task, agent)
batch = 16
prnts = 10
epis = 50000 / batch / prnts
save = False
rl = []
for updates in range(epis):